Treatment options for oncology patients are increasingly personalised with the onset of precision medicine, which enables drug selection tailored to an individual. To truly optimise the outcome, however, the treatment should also include personalised dosing (1). Faced with a trade-off between efficacy and toxicity, oncologists traditionally administer maximum tolerated dose (MTD)—the highest dose that does not cause unacceptable side effects (2), derived from Phase I trials. If patients experience toxicities, the dose is reduced according to guidelines based on population data. The key underlying assumption is that higher doses will provide greater efficacy, which many now recognize as flawed (3). Large randomised controlled trials (RCTs) show that doubling the dose of some drugs results in more frequent toxicities, which leads to dose reductions and treatment interruptions, despite no improvement in overall survival (4, 5). Conversely, other authors suggest that 30–75% of patients may be underdosed with conventional dosing strategies (6). Given the narrow therapeutic window of oncologic drugs and the 4- to 10-fold interindividual variation in pharmacokinetic clearance (7, 8), these overdosing or underdosing events result in suboptimal results: unnecessary toxicities, low efficacy and potential failure of drug development (9). The complexity of this problem increases exponentially in multi-drug regimens, with unpredictable drug-drug interactions (10, 11).

The challenge of personalised dosing has thus far been practically unsolvable because of its monumental complexity. Pharmacokinetic-based therapeutic drug monitoring and pharmacogenetic testing are promising (12–14), but have thus far not been adopted in chemotherapy drug dosing. In addition, threshold drug exposure does not necessarily correlate with threshold efficacy, particularly due to the individualised nature of patient responses to treatment. With the advent of artificial intelligence (AI), efforts are now focused on big data in pharmacogenomics (15). Though big data modelling is valuable, it is unsuitable for dynamic, multifactorial diseases, as it requires massive volumes of population information, comprehensive prior knowledge and high temporal resolution (16).

The alternative is to use small data, where only the individual's medical profile is used solely for his/her own treatment. Harnessing AI, we previously demonstrated that a quadratic surface can closely describe the relationship between drug doses and a phenotypic response in a human system (17–21). The coefficients of the second-order polynomial are unique to each individual. AI is used to map the extensive dose-response parameter space based on minimal empirical data, thus deterministically predicting the global optimum within a pre-specified safe dose range (22). Uniquely, this process is mechanism-independent and implicitly avoids the complexities stemming from the known or unknown mechanistic components, including pharmacokinetics, pharmacodynamics, pharmacogenomics, and disease biology (23, 24). These personalised profiles, which are visually represented as response surfaces, can also be scaled into higher dimensions to enable predictions for combination drug regimens (25). As such, CURATE.AI—a small data, AI-derived, indication-agnostic and mechanism-independent technology platform—could thus allow for clinically actionable personalised dosing for precision oncology and beyond. Importantly, CURATE.AI implementation is not purely computational, as it pairs prospective calibration of patient-specific responses to treatment at different drug doses with the dose optimisation process.

Predicting the optimal dose at a single point in time is only part of the challenge. The optimal dose evolves throughout the course of treatment, with the participant's response to therapy evolving due to environmental, physiologic and disease changes, including drug resistance (7). Here, CURATE.AI's ability to continually recalibrate personalised profiles allows for dynamic dose optimisation throughout treatment (23, 24). To maximise CURATE.AI's therapeutic potential, it is thus paramount to utilize a phenotypic response biomarker that can be quantified frequently. Though the gold standard for monitoring treatment response in solid tumours is the Response Evaluation Criteria in Solid Tumours (RECIST), these radiological assessments are performed at infrequent intervals after a few chemotherapy cycles, and may not reflect the entire tumour burden in situations such as widespread metastases or central necrosis (26). Blood-based tumour markers, such as carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9) (27–29), can be measured more frequently. However, these traditional blood biomarkers have limited sensitivity and specificity in reflecting tumour response, especially among patients whose baseline biomarker levels are in the normal range (30). Recent advances in genomic sequencing have revealed plasma circulating tumour DNA (ctDNA) as a novel biomarker of tumour burden (31). ctDNA are DNA fragments released by tumour cells, via apoptosis, necrosis or active secretion. These fragments contain both genetic information (such as point mutation, copy number variation) and epigenetic information (such as DNA methylation). Quantitative changes in ctDNA have been shown to be ultra-sensitive in detecting both macroscopic and microscopic tumour load (32), including in patients without raised traditional tumour markers at baseline. Therefore, serial ctDNA measurements may be more reflective of the evolving tumour response and thus appropriate for CURATE.AI.

Though CURATE.AI has already been applied in a variety of clinical settings (23–25), there are currently no prior RCTs of CURATE.AI in patients with solid tumours. Additionally, the temporal resolution of the change in the selected tumour markers, as required for CURATE.AI, has not been yet characterised. We are thus conducting this pilot clinical trial to assess the technical and logistical feasibility of a future RCT, as well as to collect exploratory data on efficacy and toxicity to enable future sample size calculations.

The inclusive recruitment criteria for this pilot clinical trial are enabled by CURATE.AI's personalisation of the treatment and permit substantial variability in the participant population that also reflect the true patient heterogeneity faced in clinical practice. This allows broader applicability of our findings and may also allow us to suggest specific subsets of patients, diseases or chemotherapy regimens where CURATE.AI-guided dosing is particularly beneficial.

Our pilot trial design has several limitations. First, this is a non-randomised pilot clinical trial, which does not simulate an RCT as closely as a randomised pilot. Therefore, it cannot inform us on feasibility issues that may arise from the logistical process or patient acceptability of randomisation, though we view such issues as unlikely in this context. Second, our recruitment criteria intentionally permit substantial variability in the participant population. This is enabled by CURATE.AI's personalisation of the treatment and reflects the true patient population heterogeneity faced in clinical practice and may allow us to identify the factors predicting successful CURATE.AI-guided dosing, for implementation in a future RCT. We nonetheless acknowledge that any identified factors will be based mainly on clinical judgement rather than statistically powered analyses. Third, participants included in this pilot must have raised levels of traditional tumour markers (e.g., CEA or CA19-9). Therefore, our experience with ctDNA in this pilot trial may not be directly generalisable to the intended patient profile in the future RCT if ctDNA were to use as an alternative biomarker for patients who lack elevated levels of traditional tumour markers. Fourth, our primary outcome measure on the successful application of CURATE.AI profiles cannot be measured based only on a numerical process and thus relies on the judgement of an expert panel with consideration for each patient's circumstances. This is inherently subjective. We have nonetheless listed the guiding criteria for the expert panel to reduce any potential bias to the minimum.

This protocol describes the design of the PRECISE CURATE.AI pilot clinical trial. To date, drug dosing in oncology lacks personalisation. CURATE.AI opens the possibility of personalised dosing for single- and multi-drug regimens, that is dynamically optimized throughout treatment. Furthermore, it is based on a small data set collected only from the treated individual rather than population data. CURATE.AI may thus overcome the challenges that impede the adoption of big data approaches for personalised drug dosing. This pilot will provide technical and logistical feasibility data, as well as exploratory efficacy and toxicity data for sample size calculations, thus laying the clinical foundation for a future RCT of personalised dosing for precision oncology using CURATE.AI.

The studies involving human participants were reviewed and approved by the National Healthcare Group (NHG) Domain Specific Review Board (DSRB), reference: 2020/00334. The patients/participants provided their written informed consent to participate in this study.

RS, DH, TK, and ABl: study conception. BT, CT, ABa, DH, RS, TK, and ABl: study design. BT, CT, XT, SP, and ABl: manuscript drafting (introduction). BT, CT, RS, XT, TK, and ABl: manuscript drafting (interventions). BT, CT, RS, TK, and ABl: manuscript drafting (outcomes). BT, CT, HS, SD, VW, RS, TK, and ABl: manuscript drafting (other methods). RS: principal investigator. All authors: critical revision.

DH, ABl, and TK are inventors of pending and issued patents pertaining to artificial intelligence-based drug development and personalised medicine. DH and TK are shareholders of KYAN Therapeutics, which has licensed intellectual property pertaining to AI-based drug development. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.